一种兼容海洋环境的改进Transformer声呐探测效能快速预报模型  

An improved Transformer model for rapid prediction of sonar detection performance compatible with marine environment

作　　者：汪晶晗 陈欢 金宇琦 兰朝凤 WANG Jinghan;CHEN Huan;JIN Yuqi;LAN Chaofeng(Hanjiang National Laboratory,Wuhan 430060,Hubei,China;School of Measurement and Communication Engineering,Harbin University of Science and Technology,Harbin 150080,Heilongjiang,China)

机构地区：[1]汉江国家实验室,武汉430060 [2]哈尔滨理工大学测控技术与通信工程学院,黑龙江哈尔滨150080

出　　处：《声学技术》2025年第2期164-170,共7页Technical Acoustics

摘　　要：为提升高复杂海洋环境下声呐探测距离预测的准确性和效率,文章提出一种基于改进Transformer的传播损失与声呐探测距离建模方法,该方法能够兼容复杂海洋环境下不同点位、不同方向声信号传播损失差异,能够基于声呐方程及声呐主被动工作模式,快速、有效地预测多点位多方向的声呐探测距离。以真实大区域海洋环境计算得到的传播损失数据为输入,通过将双向长短时记忆网络(bidirectional long short-term memory,Bi-LSTM)与Transformer架构中自注意力机制相结合,使得模型能够有效捕捉复杂环境变化的局部精确性和全局特征。实验结果表明,所提模型预测结果与声呐方程耦合积分方式得到的探测距离具有较好的一致性;同时计算效率提高了约1 000倍,提升了声呐性能的预报效率。To enhance the accuracy and efficiency of predicting sonar detection distance in high-complex marine environments,an improved Transformer-based modeling method for transmission loss and sonar detection distance is proposed in this paper.This method can accommodate the differences in transmission losses across various positions and directions in complex marine environments,and can precisely and rapidly predict multi-point and multi-directional sonar detection distances based on sonar equations and active/passive operating modes.Taking the transmission loss data calculated in the real large regional marine environment as input,by incorporating bidirectional long short-term memory(Bi-LSTM)network with the self-attention mechanism of the Transformer architecture,the proposed model is able to accurately capture both local details and global features in response to environmental variations.Experimental results show that the outcomes predicted from this model exhibit good consistency with the detection radii derived from the sonar equation coupling integration method.Additionally,the computational efficiency is improved by approximately 1000-fold,significantly improving the efficiency of predicting sonar performance.

关 键 词：声呐性能快速预测 深度学习 双向长短时记忆网络(Bi-LSTM) Transformer架构 

分 类 号：TB56[交通运输工程—水声工程]